Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
uiobmcrop
STK3405 – Week 39
A. B. Huseby & K. R. Dahl
Department of MathematicsUniversity of Oslo, Norway
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 1 / 37
uiobmcrop
Chapter 5
Structural and reliability importance for components inbinary monotone systems
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 2 / 37
uiobmcrop
Importance measures
A measure of importance can be used to identify components thatshould be improved in order to increase the system reliability.
A measure of importance can be used to identify components thatmost likely have failed, given that the system has failed.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 3 / 37
uiobmcrop
Section 5.1
Structural importance of a component
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 4 / 37
uiobmcrop
Criticality
Definition (Criticality)
Let (C, φ) be a binary monotone system, and let i ∈ C. We say thatcomponent i is critical for the system if:
φ(1i ,x) = 1 and φ(0i ,x) = 0.
If this is the case, we also say that (·i ,x) is a critical vector forcomponent i.
NOTE: Criticality is strongly related to the notion of relevance: Acomponent i in a binary monotone system (C, φ) is relevant if and onlyif there exists at least one critical vector for i .
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 5 / 37
uiobmcrop
Criticality (cont.)
2 3 4
1
Figure: A binary monotone system (C, φ)
The structure function of the system (C, φ) is given by:
φ(x) = x1 q (x2 · x3 · x4)
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 6 / 37
uiobmcrop
Criticality (cont.)
Component 1 is critical if (·1,x) is:
(·,0,0,0), (·,1,0,0), (·,0,1,0), (·,0,0,1),
(·,1,1,0), (·,1,0,1), (·,0,1,1).
Component 2 is critical if (·2,x) = (0, ·,1,1),
Component 3 is critical if (·3,x) = (0,1, ·,1),
Component 4 is critical if (·4,x) = (0,1,1, ·).
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 7 / 37
uiobmcrop
Structural importance
Based on this Birnbaum suggested the following measure of structuralimportance of a component in a binary monotone system:
Definition (Structural importance)
Let (C, φ) be a binary monotone system of order n, and let i ∈ C. TheBirnbaum measure for the structural importance of component i, denoted J(i)
B ,is defined as:
J(i)B :=
12n−1
∑(·i ,x )
[φ(1i ,x)− φ(0i ,x)].
Note that the denominator, 2n−1 is the total number of states for the n − 1other components. Thus, J(i)
B can be interpreted as the fraction of all statesfor the n − 1 other components where component i is critical.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 8 / 37
uiobmcrop
Structural importance
2 3 4
1
Figure: A binary monotone system (C, φ)
For this system we have the following structural importance measures:
J(1)B =
724−1 =
78, J(2)
B = J(3)B = J(4)
B =1
24−1 =18.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 9 / 37
uiobmcrop
Structural importance
Let φ be a 2-out-of-3 system. To compute the structural importance ofcomponent 1, we note that the critical vectors for this component are(·,1,0) and (·,0,1). Hence, we have:
J(1)B =
223−1 =
12.
By similar arguments, we find that:
J(2)B = J(3)
B =12.
So in a 2-out-of-3 system, all of the components have the samestructural importance. This is intuitively obvious since the structurefunction is symmetrical with respect to the components.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 10 / 37
uiobmcrop
Section 5.2
Reliability importance of a component
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 11 / 37
uiobmcrop
Reliability importance of a component
Definition (Reliability importance of a component)
Let (C, φ) be a binary monotone system, and let i ∈ C. Moreover, let Xbe the vector of component state variables.
The Birnbaum measure for the reliability importance of component i,denoted I(i)B is defined as:
I(i)B := P(Component i is critical for the system)
= P(φ(1i ,X )− φ(0i ,X ) = 1).
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 12 / 37
uiobmcrop
Reliability importance of a component (cont.)
Since the difference φ(1i ,X )− φ(0i ,X ) is a binary variable, it followsthat:
I(i)B = E [φ(1i ,X )− φ(0i ,X )] = E [φ(1i ,X )]− E [φ(0i ,X )].
In particular, if the component state variables of the system areindependent, and P(Xi = 1) = pi for i ∈ C, we get that:
I(i)B = h(1i ,p)− h(0i ,p).
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 13 / 37
uiobmcrop
Reliability importance of a component (cont.)
Theorem (Partial derivative formula)
Let (C, φ) be a binary monotone system where the component state variablesare independent, and P(Xi = 1) = pi for i ∈ C.
Then:I(i)B =
∂h(p)∂pi
, for all i ∈ C.
PROOF: By pivotal decomposition we have:
h(p) = pih(1i ,p) + (1− pi)h(0i ,p)
By differentiating this identity with respect to pi we get:
∂h(p)∂pi
= h(1i ,p)− h(0i ,p).
Hence, the result follows.A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 14 / 37
uiobmcrop
Reliability importance inequalities
Theorem (Reliability importance inequalities)
For a binary monotone system, (C, φ), we always have
0 ≤ I(i)B ≤ 1.
Assume that the component state variables are independent, andP(Xj = 1) = pj , where 0 < pj < 1 for all j ∈ C.
If component i is relevant, we have:
0 < I(i)B .
Furthermore, if there exists at least one other relevant component, wealso have:
I(i)B < 1.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 15 / 37
uiobmcrop
Reliability importance inequalities (cont.)
PROOF: We note that the first inequality follows directly from thedefinition since the reliability importance is a probability.
We then assume that the component state variables are independent,and that P(Xj = 1) = pj , where 0 < pj < 1 for all j ∈ C.
If component i is relevant, we know that h is strictly increasing in pi .
That is, we must have:∂h(p)∂pi
> 0.
Combining this with the partial derivative formula, we get that 0 < I(i)B .
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 16 / 37
uiobmcrop
Reliability importance inequalities (cont.)
Finally, we assume that there exists at least one other relevantcomponent k ∈ C.
To show that this implies that I(i)B < 1, we assume instead that I(i)B = 1,and show that this leads to a contradiction.
By this assumption, it follows that :
P(φ(1i ,X )− φ(0i ,X ) = 1) = 1
Since 0 < pj < 1, for all j ∈ C, it follows that P((·i ,X ) = (·i ,x)) > 0 forall (·i ,x).
Hence, we must have that:
φ(1i ,x) = 1 and φ(0i ,x) = 0 for all (·i ,x).
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 17 / 37
uiobmcrop
Reliability importance inequalities (cont.)
At the same time, since component k is relevant, there exists a vector(·k ,y) such that:
φ(1k ,y) = 1 and φ(0k ,y) = 0.
If yi = 1, it follows that φ(1i ,0k ,y) = 0, contradicting that φ(1i ,x) = 1for all (·i ,x).
If yi = 0, it follows that φ(0i ,1k ,y) = 1, contradicting that φ(0i ,x) = 0for all (·i ,x).
Hence, we conclude that for both possible values of yi we end up withcontradictions.
Thus, the only possibility is that I(i)B < 1.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 18 / 37
uiobmcrop
Reliability importance and structural importance
Theorem (Reliability importance and structural importance)
Consider a binary monotone system (C, φ) where the component statevariables are independent, and where P(Xi = 1) = 1
2 for all i ∈ C.Then we have:
I(i)B = J(i)B
PROOF: If the component state variables are independent, andP(Xi = 1) = 1
2 for all i ∈ C, we have:
P((·i ,X ) = (·i ,x)) =∏j 6=i
P(Xj = xj) =∏j 6=i
(12) =
12n−1 .
From this the result follows.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 19 / 37
uiobmcrop
Reliability importance examples
In the following examples we consider binary monotone systems (C, φ)where C = {1, . . . ,n}.
We also assume that the component state variables are independent,and that:
P(Xi = 1) = pi , i ∈ C.
Without loss of generality we assume that the components are orderedso that:
p1 ≤ p2 ≤ . . . ≤ pn. (1)
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 20 / 37
uiobmcrop
Reliability importance examples (cont.)
Let (C, φ) be a series system. Then for all i ∈ C we have:
I(i)B =∂∏n
j=1 pj
∂pi=
∏j 6=i
pj .
Hence, by the ordering (1), we get that:
I(1)B ≥ I(2)B ≥ · · · ≥ I(n)B .
Thus, in a series system the worst component, i.e., the one with thesmallest reliability, has the greatest reliability importance.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 21 / 37
uiobmcrop
Reliability importance examples (cont.)
Let (C, φ) be a parallel system. Then for all i ∈ C we have:
I(i)B =∂∐n
j=1 pj
∂pi=∂[1−
∏nj=1(1− pj)]
∂pi=
∏j 6=i
(1− pj).
Hence, from the ordering (1)
I(1)B ≤ I(2)B ≤ · · · ≤ I(n)B .
Thus, in a parallel system the best component, i.e., the one with thegreatest reliability, has the greatest reliability importance.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 22 / 37
uiobmcrop
Reliability importance examples (cont.)
Let (C, φ) be a 2-out-of-3 system. It is then easy to show that:
φ(X ) = X1X2 + X1X3 + X2X3 − 2X1X2X3.
Hence, we have:
h(p) = p1p2 + p1p3 + p2p3 − 2p1p2p3.
This implies that:
I(1)B =∂h(p)∂p1
= p2 + p3 − 2p2p3,
I(2)B =∂h(p)∂p2
= p1 + p3 − 2p1p3,
I(3)B =∂h(p)∂p3
= p1 + p2 − 2p1p2.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 23 / 37
uiobmcrop
Reliability importance examples (cont.)
We then consider the function f (p,q) = p + q − 2pq and note that:
I(1)B = f (p2,p3), I(2)B = f (p1,p3), I(3)B = f (p1,p2).
Moreover, the partial derivatives of f are respectively:
∂f∂p
= 1− 2q,∂f∂q
= 1− 2p.
If p,q ≤ 12 , f is non-decreasing in p and q. Thus, if p1 ≤ p2 ≤ p3 ≤ 1
2 , wehave:
f (p1,p2) ≤ f (p1,p3) ≤ f (p2,p3).
Hence, in this case we have:
I(3)B ≤ I(2)B ≤ I(1)B . (2)
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 24 / 37
uiobmcrop
Reliability importance examples (cont.)
If p,q ≥ 12 , f is non-increasing in p and q. Thus, if 1
2 ≤ p1 ≤ p2 ≤ p3, wehave:
f (p2,p3) ≤ f (p1,p3) ≤ f (p1,p2).
Hence, in this case we have:
I(1)B ≤ I(2)B ≤ I(3)B . (3)
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 25 / 37
uiobmcrop
Reliability importance examples (cont.)
If p1 = 12 − z, p2 = 1
2 and p3 = 12 + z, where z ∈ (0, 1
2 ), we get:
I(1)B = (12) + (
12+ z)− 2 · (1
2)(
12+ z) =
12,
I(2)B = (12− z) + (
12+ z)− 2 · (1
2− z)(
12+ z) =
12+ 2z2,
I(3)B = (12− z) + (
12)− 2 · (1
2− z)(
12) =
12,
Hence in this case we have:
I(1)B = I(3)B ≤ I(2)B . (4)
Note that this result holds also if z ∈ (− 12 ,0) in which case p1 > p2 > p3.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 26 / 37
uiobmcrop
Reliability importance examples (cont.)
2 3 4
1
Figure: A binary monotone system (C, φ)
The structure function of this system is:
φ(X ) = X1 q (X2 · X3 · X4) = X1 + X2 · X3 · X4 − X1 · X2 · X3 · X4
Thus, the reliability function is given by:
h(p) = p1 + p2 · p3 · p4 − p1 · p2 · p3 · p4
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 27 / 37
uiobmcrop
Reliability importance examples (cont.)
Hence we have:
I(1)B = 1− p2 · p3 · p4
I(2)B = p3 · p4 − p1 · p3 · p4 = (1− p1) · p3 · p4
I(3)B = p2 · p4 − p1 · p2 · p4 = (1− p1) · p2 · p4
I(4)B = p2 · p3 − p1 · p2 · p3 = (1− p1) · p2 · p3
If p1 = p2 = p3 = p4 = p ∈ (0,1), we have:
I(1)B = 1− p3
I(i)B = p2 − p3 < I(1)B , i = 2,3,4.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 28 / 37
uiobmcrop
Reliability importance examples (cont.)
Assume instead that p1 = 0.1 and that p2 = p3 = p4 = 0.9. Then weget:
I(1)B = 1− p2 · p3 · p4 = 1− 0.93 = 0.271
I(2)B = p3 · p4 − p1 · p3 · p4 = (1− p1) · p3 · p4 = 0.93 = 0.729
I(3)B = p2 · p4 − p1 · p2 · p4 = (1− p1) · p2 · p4 = 0.93 = 0.729
I(4)B = p2 · p3 − p1 · p2 · p3 = (1− p1) · p2 · p3 = 0.93 = 0.729
Thus, in this case we have:
I(1)B < I(2)B = I(3)B = I(4)B .
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 29 / 37
uiobmcrop
Section 5.3
The Barlow-Proschan measure of reliability importance
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 30 / 37
uiobmcrop
The Barlow-Proschan measure of reliabilityimportance
Definition (Barlow-Proschan measure)
Consider a binary monotone system (C, φ), where the components arenever repaired.
Moreover, let Ti denote the lifetime of component i, i ∈ C, and let Sdenote the lifetime of the system.
The Barlow-Proschan measure of the reliability importance ofcomponent i ∈ C is defined as:
I(i)B−P := P(Component i fails at the same time as the system)
= P(Ti = S).
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 31 / 37
uiobmcrop
Absolute continuity
A real-valued stochastic variable, T has an absolutely continuousdistribution if P(T ∈ A) = 0 for all measurable sets A ⊆ R such thatm1(A) = 0, where m1 denotes the Lebesgue measure in R.
If T1, . . . ,Tn are independent and absolutely continuously distributed,then T = (T1, . . . ,Tn) is absolutely continuously distributed in Rn.
That is, P(T ∈ A) = 0 for all (measurable) sets A ⊆ Rn such thatmn(A) = 0, where mn denotes the Lebesgue measure in Rn.
In particular, if A = {t : ti = tj}, where i 6= j , then mn(A) = 0.
Hence, P(Ti = Tj) = 0 when i 6= j .
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 32 / 37
uiobmcrop
The Barlow-Proschan measure of reliabilityimportance (cont.)
Theorem (Probability of system failure)
Let (C, φ) be a non-trivial binary monotone system where thecomponents are never repaired and C = {1, . . . ,n}.Let Ti denote the lifetime of component i, i = 1, . . . ,n, and let S denotethe lifetime of the system.
Moreover, assume that T1, . . . ,Tn are independent, absolutelycontinuously distributed.
Then S is absolutely continuously distributed as well, and we have:
n∑i=1
I(i)B−P = 1.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 33 / 37
uiobmcrop
The Barlow-Proschan measure of reliabilityimportance (cont.)
PROOF: Since we have assumed that the system is non-trivial, thelifetime of the system, S can be expressed as:
S = max1≤j≤p
mini∈Pj
Ti , (5)
where P1, . . . ,Pp are the minimal path sets of the system. This impliesthat:
P(n⋃
i=1
{Ti = S}) = 1. (6)
Let A ⊆ R be an arbitrary measurable set such that m1(A) = 0. Sincewe have assumed that T1, . . . ,Tn are absolutely coninuouslydistributed, we get that:
0 ≤ P(S ∈ A) ≤ P(n⋃
i=1
{Ti ∈ A}) ≤n∑
i=1
P(Ti ∈ A) = 0,
which implies that S is absolutely continuously distributed as well.A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 34 / 37
uiobmcrop
The Barlow-Proschan measure of reliabilityimportance (cont.)
Since T1, . . . ,Tn are absolutely continuously distributed, the probability ofhaving two or more components failing at the same time is zero.
This implies e.g., that P({Ti = S} ∩ {Tj = S}) = 0 for i 6= j . Thus, whencalculating the probability of the union of the events {Ti = S}, i = 1, . . . ,n, allintersections can be ignored as they have zero probability of occurring.
Hence, by (6) we get:
1 = P(n⋃
i=1
{Ti = S}) =n∑
i=1
P(Ti = S) =n∑
i=1
I(i)B−P ,
where the second equality follows by ignoring all intersections of events{Ti = S}, i = 1, . . . ,n.
The last equality follows by the definition of I(i)B−P , and hence, the proof iscomplete.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 35 / 37
uiobmcrop
The Barlow-Proschan measure of reliabilityimportance (cont.)
Theorem (Integral formula for the Barlow-Proschan measure)
Let (C, φ) be a non-trivial binary monotone system where the componentsare never repaired, and where C = {1, . . . ,n}Let Ti denote the lifetime of component i, i = 1, . . . ,n.
Moreover, assume that T1, . . . ,Tn are independent, absolutely continuouslydistributed with densities f1, . . . , fn respectively.
Then, we have:
I(i)B−P =
∫ ∞0
I(i)B (t)fi(t)dt ,
where I(i)B (t) denotes the Birnbaum measure of the reliability importance ofcomponent 1 at time t, i.e., the probability that component i is critical at time t.
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 36 / 37
uiobmcrop
The Barlow-Proschan measure of reliabilityimportance (cont.)
PROOF: From the definitions of the Barlow-Proschan measure and theBirnbaum measure, it follows that:
I(i)B−P = P(Component i fails at the same time as the system)
=
∫ ∞0
P(Component i is critical at time t) · fi(t)dt
=
∫ ∞0
I(i)B (t)fi(t)dt .
A. B. Huseby & K. R. Dahl (Univ. of Oslo) STK3405 – Week 39 37 / 37